WO2022119256A1

WO2022119256A1 - Drug response prediction device using self-attention-based hierachical network, and method therefor

Info

Publication number: WO2022119256A1
Application number: PCT/KR2021/017761
Authority: WO
Inventors: 남호정; 진일중
Original assignee: 광주과학기술원
Priority date: 2020-12-01
Filing date: 2021-11-29
Publication date: 2022-06-09
Also published as: KR20220077111A

Abstract

The present invention relates to an in-silico type drug sensitivity prediction device and method using a hierarchical artificial intelligence network. According to the present invention, by predicting drug sensitivity by using a self-attention-based artificial intelligence network, changes in cell state such as changes in gene expression level can be reflected to the network, and thus, it is possible to more accurately predict a drug response.

Description

Drug response prediction device and method using self-attention-based hierarchical network

The present invention relates to drug reactivity prediction, and more particularly, to drug reactivity prediction technology using an artificial neural network.

In the biotechnology research method, the method of experimenting based on living organisms is called in vivo, and the method through a glass test tube is called in-vitro.

In case of testing drug response by culturing cells in laboratory animals or test tubes, it faces ethical issues as well as time and cost issues. In-silico methods are being tried.

In particular, with the development of artificial intelligence learning methods such as deep learning and machine learning, various attempts are being made to find drugs suitable for cancer cell lines by learning drug responsiveness to artificial intelligence. Learning of artificial intelligence can be accomplished using a database called GDSC (Genomics of Drug Sensitivity in Cancer).

However, it is not easy to find drug responsiveness to new cancer cells using the learned artificial intelligence. This is because the learned network of artificial intelligence may not be suitable for new cancer cells, and the importance of a gene may vary depending on the expression level of the gene because the state of the cell continues to change even for cancer cells that have already been learned.

The inventors of the present invention have been working hard to solve the problems of drug reactivity prediction using artificial intelligence of the prior art. After much effort, the present invention was completed after much effort to complete a drug responsiveness prediction device and method including an artificial intelligence network that can reflect the importance that varies depending on the expression level of genes for new or existing cancer cells.

An object of the present invention is to provide an apparatus and method for predicting drug reactivity including an artificial intelligence network that can reflect the state of cells, such as the amount of gene expression in cells that change frequently based on the self-attention technique. .

On the other hand, other objects not specified in the present invention will be additionally considered within the range that can be easily inferred from the following detailed description and effects thereof.

Self-attention-based drug reactivity prediction device according to the present invention,

a gene-level network that outputs pathway activity in a gene-level self-attention-based artificial intelligence network that inputs gene expression levels and drug descriptors for each pathway; a pathway-level network that receives the pathway activity as an input and outputs a pathway feature multiplied by a weight to a pathway-level self-attention-based AI network; a drug network that outputs drug descriptors using a drug deep learning network (DNN); and a reactivity prediction network for outputting a drug response prediction using a drug response deep learning network by inputting the weighted pathway feature and the drug descriptor as inputs.

The gene-level network obtains a gene-level attention score by using a value passed through a gene-level weight network with the gene expression level and drug descriptor as inputs, and the gene-level attention score and the gene expression level It is characterized in that the pathway activity is obtained by a value multiplied by .

The pathway level network obtains a pathway level attention score as a value passed through a pathway level weight network with the pathway activity and drug descriptor as inputs, and multiplies the pathway level attention score by the pathway activity. It is characterized in that a path feature multiplied by a weight is obtained.

The drug descriptor is characterized in that it is obtained using a Morgan fingerprint.

The reactivity prediction network is characterized in that the drug response prediction is output as an IC50 predicted value.

A self-attention-based drug reactivity prediction method according to another embodiment of the present invention,

(a) outputting a pathway activity in a gene-level self-attention-based artificial intelligence network with gene expression levels and drug descriptors for each pathway as inputs; (b) outputting a weighted pathway feature to a pathway-level self-attention-based AI network using the pathway activity as an input; (c) outputting a drug descriptor using a drug deep learning network (DNN); and (d) outputting a drug response prediction using a drug response deep learning network by inputting the weighted pathway feature and the drug descriptor as inputs.

The step (a) is to obtain a gene level attention score (Attention Score) using the value passed through the gene level weight network as the input of the gene expression level and the drug descriptor, the gene level attention score and the gene expression It is characterized in that the pathway activity is calculated as a value multiplied by the amount.

In the step (b), the pathway-level network obtains a pathway-level attention score using a value passed through a pathway-level weighting network with the pathway activity and drug descriptor as inputs, and the pathway-level attention score and the pathway A value multiplied by the activity is characterized in that a pathway feature multiplied by the weight is obtained.

The drug descriptor in step (c) is characterized in that it is obtained using a Morgan fingerprint.

The step (d) is characterized in that the drug response prediction is output as an IC50 predicted value.

According to the present invention, since a hierarchical network is used based on the self-attention technique, even if the cell state continues to change, such as a change in gene expression level, the network state is changed to reflect this, and it is possible to predict an accurate drug response.

On the other hand, even if it is an effect not explicitly mentioned herein, it is added that the effects described in the following specification expected by the technical features of the present invention and their potential effects are treated as described in the specification of the present invention.

1 is a schematic structural diagram of a self-attention-based drug reactivity prediction device according to a preferred embodiment of the present invention.

2 is a schematic structural diagram of a gene level network according to a preferred embodiment of the present invention.

3 is a schematic structural diagram of a pathway level network according to a preferred embodiment of the present invention.

4 is a schematic structural diagram of a drug network according to a preferred embodiment of the present invention.

5 is a schematic structural diagram of a reactive prediction network according to a preferred embodiment of the present invention.

6 is a graph showing the drug reactivity results predicted by the self-attention-based drug reactivity prediction device according to a preferred embodiment of the present invention.

7 is a flowchart of a method for predicting self-attention-based drug reactivity according to another preferred embodiment of the present invention.

※ It is revealed that the accompanying drawings are exemplified as a reference for understanding the technical idea of the present invention, and the scope of the present invention is not limited thereby

Hereinafter, the configuration of the present invention guided by various embodiments of the present invention and effects resulting from the configuration will be described with reference to the drawings. In the description of the present invention, if it is determined that related known functions are obvious to those skilled in the art and may unnecessarily obscure the gist of the present invention, the detailed description thereof will be omitted.

Terms such as 'first' and 'second' may be used to describe various elements, but the elements should not be limited by the above terms. The above term may be used only for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, a 'first component' may be termed a 'second component', and similarly, a 'second component' may also be termed a 'first component'. can Also, the singular expression includes the plural expression unless the context clearly dictates otherwise. Unless otherwise defined, terms used in the embodiments of the present invention may be interpreted as meanings commonly known to those of ordinary skill in the art.

Hereinafter, the configuration of the present invention guided by various embodiments of the present invention and effects resulting from the configuration will be described with reference to the drawings.

Self-attention-based drug reactivity prediction apparatus 100 according to the present invention is a gene-level network 110 , a pathway-level network 120 , a drug network 130 , and reactivity and a Response Prediction network 140 .

The self-attention-based drug reactivity prediction apparatus 100 predicts the reactivity between a drug and cancer cells using an artificial intelligence network. For realizing an artificial intelligence network, the self-attention-based drug reactivity prediction apparatus 100 may include a processor and a memory, and program codes for driving the processor and learned network data may be stored in the memory.

When a network weight is learned in an artificial neural network of the prior art, the reactivity between a drug and a gene or pathway is predicted by a fixed weight. By constructing an artificial intelligence neural network based on the Self-Attention technique to obtain the Attention Score, it has the advantage of being able to check the change in importance according to the state of each cell or the amount of gene expression.

To this end, the gene-level network 110 and the pathway-level network 120 of the present invention use self-attention-based artificial intelligence neural network networks.

The gene-level network 110 receives the gene expression level 10 and the drug descriptor 20 of the sample and outputs pathway activity based on self-attention, and the pathway level (Pathway- level) the network 120 receives the pathway activity as an input and outputs the pathway activity multiplied by the weight.

The reactivity prediction network 140 receives the pathway activity multiplied by the weight from the pathway level network 120 and the drug network 130 and the drug descriptor multiplied by the weight, and outputs an IC50 value 30 . IC50 refers to the amount of drug that can eliminate half (50%) of cancer cells. The drug descriptor uses the Morgan Fingerprint.

Using this self-attention-based network, it is possible to specify a gene or pathway that has a high attention score, that is, a target gene or pathway for a drug, rather than simply how effective a drug is on which cancer cell. That is, unlike the prior art, an interpretable network can be used.

2 to 5 show the structure of each network in more detail.

The gene level network 110 receives the gene expression level 10 and the drug descriptor 20 of the sample as inputs and outputs the pathway activity 12 based on self-attention.

First, an output obtained by multiplying the gene expression level (10) and the drug descriptor (20) by the weight of the network is obtained. This is expressed by the following equation.

u _ij = tanh(W _G [gd] + b _G ) _ij

When the gene expression level (g) and the drug descriptor (d) are multiplied by the network weight (W _G ) and the bias (b _G ) is added to pass through the network, the network output (u _ij ) is obtained as an output.

If the soft max value of this network output (u _ij ) is obtained, the attention score (α _ij ) can be obtained as follows.

This attention score (α _ij ) is multiplied by the gene expression level (g _ij ) again to obtain the pathway activity (12, p _i ) for each pathway as follows.

With the previously obtained pathway activity 12 as an input, the pathway activity 14 multiplied by a weight based on self-attention is obtained.

First, the pathway activity 12 and the drug descriptor 20 are input to the network, and the network output value u _i is obtained as follows.

u _i = tanh(W _P [pd] + b _p ) _i

By calculating the soft max value for the network output value (u _i ), the attention score (α _i ) can be obtained, which is expressed by the following equation.

If this attention score (α _i ) and the previously obtained pathway activity 12 are passed through the network again, the weighted pathway activity 14, s can be obtained.

In this way, the gene level network 110 and the pathway level network 120 do not simply pass the input value through the artificial intelligence network, but continuously reflect the importance to the network, so that the network changes depending on the cell or the state of the cell and can be applied. can have an effect.

The drug network 130 outputs the drug descriptor 22 passing the drug descriptor 20 through a general artificial neural network. A deep neural network (DNN) or the like may be used as the artificial neural network. The drug descriptors 22 and c that have passed through the drug network 130 may be expressed by the following equation.

c = ReLU(W _D d + b _D )

Finally, when the weighted pathway 14 and the drug descriptor 22 that have passed through the artificial neural network are connected and pass through an artificial neural network such as DNN, an IC50 value 30 of the drug for the sample is finally obtained.

The drug response value (r) can be expressed by the following formula.

r = ReLU(W _R [sc]+b _R )

Using this self-attention-based attention score, it is possible not only to judge whether a drug is effective against which cancer cells, but also to translate ( Interpretation) is possible.

6a shows a comparison of predicted values and observed values for a case where both the cell line and the drug are unknown (Unseen Pair), and b is the cell line when the cell line is unknown. It shows the comparison of the predicted value and the observed value at the time (Unseen Cell Line). It shows that when the IC50 is predicted based on self-interest, an improved result is obtained compared to the prior art.

The self-attention-based drug reactivity prediction method according to the present invention may be performed by a controller including one or more processors and a memory.

Pathway activity is first calculated by the network learned using a database such as GDSC (S10).

The gene expression level of the sample and the drug descriptor are passed through the network as inputs, the attention score is calculated based on the output value, and then it is multiplied by the gene expression level to calculate the pathway activity.

If the pathway activity is obtained, it is again input to the network together with the drug descriptor to obtain an output value, and by multiplying the attention score by the output value by the pathway activity, a pathway feature that is a weighted pathway activity can be obtained (S20).

The drug descriptor is calculated by a method such as Morgan fingerprint, and the output is obtained through an artificial neural network such as DNN (S30). can be (S40).

According to the self-attention-based drug response prediction device and method as described above, it is possible to predict more accurately drug response according to the cell state or gene expression level by constructing a network using the attention score that reflects importance rather than a fixed network weight. It works.

The protection scope of the present invention is not limited to the description and expression of the embodiments explicitly described above. In addition, it is added once again that the protection scope of the present invention cannot be limited due to obvious changes or substitutions in the technical field to which the present invention pertains.

The self-attention-based drug reactivity prediction device according to the present invention is a drug reactivity prediction technology using an artificial neural network and can be applied to medicine and medical fields.

Claims

In the self-attention-based drug reactivity prediction device comprising a controller including one or more processors and a memory, the controller comprises:

Gene-level self-attention-based artificial intelligence neural network that inputs gene expression level and drug descriptor by pathway for arbitrary cell lines and drugs Gene level that outputs pathway activity network;

a pathway-level network that receives the pathway activity as an input and outputs a pathway feature multiplied by a weight to a pathway-level self-attention-based artificial intelligence neural network network;

a drug network that outputs drug descriptors using a drug deep learning network (DNN); and

Responsiveness prediction network for outputting drug response prediction using the drug response deep learning network by inputting the weighted pathway feature and the drug descriptor as inputs;

The gene level network obtains a gene level attention score by using a value passed through a gene level weight network with the gene expression level and drug descriptor as inputs, and multiplies the gene level attention score with the gene expression level to pass the pass Find the way activity,

The pathway level network is configured to obtain a pathway level attention score as a value passed through a pathway level weight network with the pathway activity and drug descriptors as inputs, and multiply the pathway level attention score by the pathway activity to pass the path Characterized in finding a way feature,

The gene level attention score is applied with a weight modified by the input of the gene expression level and drug descriptor,

The pathway level attention score is applied with a weight modified by the input of the pathway activity and drug descriptor,

the control unit

By determining the importance of the gene and pathway based on the gene-level attention score and the pathway-level attention score, specifying the gene and pathway that are the target of the drug

Self-Awareness-Based Drug Reactivity Prediction Device.
According to claim 1,

The drug descriptor is characterized in that it is obtained using a Morgan fingerprint (Morgan Fingerprint), self-attention-based drug reactivity prediction device.
According to claim 1,

The self-attention-based drug reactivity prediction device, characterized in that the reactivity prediction network outputs the drug response prediction as an IC50 predicted value.
In the self-attention-based drug reactivity prediction method performed by a control unit including one or more processors and memory:

(a) Pathway activity is output from a gene-level self-attention-based artificial intelligence neural network that inputs gene expression levels and drug descriptors by pathway for arbitrary cell lines and drugs to do;

(b) outputting a weighted pathway feature to a pathway-level self-attention-based artificial intelligence neural network using the pathway activity as an input;

(c) outputting a drug descriptor using a drug deep learning network (DNN); and

(d) outputting a drug response prediction using a drug response deep learning network by inputting the weighted pathway feature and the drug descriptor as inputs;

In step (a), a gene level attention score is obtained using a value passed through a gene level weight network as an input with the gene expression level and drug descriptor, and the gene level attention score is multiplied by the gene expression level. Find the pathway activity,

In the step (b), a pathway level attention score is obtained as a value passed through a pathway level weight network with the pathway activity and drug descriptors as inputs, and the pathway level attention score is multiplied by the pathway activity to pass the path Characterized in finding a way feature,

The step (a) determines the importance of the gene based on the gene-level attention score to specify the gene as a target of the drug, and the step (b) determines the pass-through level based on the pathway level attention score determining the importance of the way and specifying the pathway that is the target of the drug,

The gene level attention score is applied with a weight modified by the input of the gene expression level and drug descriptor,

The pathway level attention score is a weighted value modified by the input of the pathway activity and drug descriptor.

A method for predicting drug reactivity based on self-attention.
5. The method of claim 4,

In the step (c), the drug descriptor is obtained by using a Morgan fingerprint, a self-attention-based drug reactivity prediction method.
5. The method of claim 4,

In the step (d), the self-attention-based drug reactivity prediction method, characterized in that the drug response prediction is output as an IC50 predicted value.